- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi,
I was trying to do INT8 optimization on a tensorflow model .
Model Name- faster_rcnn_inception_v2_coco_2018_01_28
Initially created OpenVINO IR files using below command:
python mo_tf.py --input_model frozen_inference_graph.pb --output_dir IR-FP32 --data_type FP32 --tensorflow_use_custom_operations_config openvino_2021.1.110/deployment_tools/model_optimizer/extensions/front/tf/faster_rcnn_support_api_v1.14.json --tensorflow_object_detection_api_pipeline_config pipeline.config --reverse_input_channels --input_shape [1,450,450,3]
Which created .xml and .bin files with the below 2 warnings:
[ WARNING ] Model Optimizer removes pre-processing block of the model which resizes image keeping aspect ratio. The Inference Engine does not support dynamic image size so the Intermediate Representation file is generated with the input image size of a fixed size.
The Preprocessor block has been removed. Only nodes performing mean value subtraction and scaling (if applicable) are kept.
The graph output nodes "num_detections", "detection_boxes", "detection_classes", "detection_scores" have been replaced with a single layer of type "Detection Output". Refer to IR catalogue in the documentation for information about this layer.
[ WARNING ] Network has 2 inputs overall, but only 1 of them are suitable for input channels reversing.
Suitable for input channel reversing inputs are 4-dimensional with 3 channels
All inputs: {'image_tensor': [1, 3, 450, 450], 'image_info': [1, 3]}
Suitable inputs {'image_tensor': [1, 3, 450, 450]}
Now, I need to convert this fp32 models to int8. I tried with faster_rcnn_resnet50_coco_int8.json and faster_rcnn_resnet50_coco_int8.yml
But, it gives the below error :
return data.reshape(input_shape) if not self.disable_resize_to_input else data
ValueError: cannot reshape array of size 1843200 into shape (1,3,450,450)
Could you please help me to do the optimization.
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Greetings,
You can refer here to Calibrate your FP32 model to INT8: https://www.youtube.com/watch?v=XkD8ae8uWes
and also the scale values are really important and they need to be precise according to your model. This is how you can get the exact scale values: https://www.youtube.com/watch?v=-8_yRzN-fTY
Sincerely,
Iffa
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Could you please suggest with an example of json and yml file to be used while doing OpenVINO
Post-Training Optimization on my specified openvino FP32 faster_rcnn_inception_v2_coco_2018_01_28 model?
- Tags:
- Could
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
This might help:
https://docs.openvinotoolkit.org/2021.2/workbench_docs_Workbench_DG_Import_TensorFlow.html
Check on the other tabs too.
Sincerely,
Iffa
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
You have shared the link to convert .pb to fp32/fp16, which I have done already.
I need the help to do INT8 conversion. Could you please help to do convert the FP32 faster_rcnn_inception_v2_coco_2018_01_28 model to int8?
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
If possible can you share your model here?
Sincerely,
Iffa
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Please find the model in the below link:
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
We are investigating this and will get back to you asap.
Sincerely,
iffa
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Greetings,
We were able to quantize the FP32 IR model.
Note: If you are using Linux use python3 instead of python command.
First, use model downloader to download the model: ( you should find these in deployment_tools/tools/model_downloader)
python downloader.py --name faster_rcnn_inception_v2-coco
Then convert using converter.py:
python converter.py --name faster_rcnn_inception_v2-coco
if this doesn't work use :
python converter.py --name faster_rcnn_inception_v2-coco --mo <location of model_optimizer> --precision FP32
Next, you need to download and cut the
coco dataset:
I attached the files that you need to use. You may refer here to perform the POT: https://docs.openvinotoolkit.org/latest/pot_configs_examples_README.html
Then you should have the INT8 file.
Sincerely,
Iffa
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Greetings,
Intel will no longer monitor this thread since we have provided a solution. If you need any additional information from Intel, please submit a new question.
Sincerely,
Iffa

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page